search

cailab-tamu / scTenifoldKnk

R/MATLAB package to perform virtual knockout experiments on single-cell gene regulatory networks.
36 stars 6 forks source link

Regarding Log2FC value #23

Open lagom2728 opened 1 year ago

lagom2728 commented 1 year ago

Dear developers

Thank you for developing this wonderful tool. Thanks to you, we were able to obtain interesting results using the tool you developed. However, I have a question regarding FC (Foldchange), the result obtained after Virtual KO analysis.

Is it theoretically okay to transform the FC value into log2FC by adding log2 ((ex) log2FC <- log2(FC))? If it is okay to transform the value into log2FC, is it okay to classify it as an up-regulated gene or a down-regulated gene?

(ex) log2(FC) > 1 = Up regulatory genes log2(FC) < -1 = Down regulatory genes

Thank you

Rohit-Satyam commented 10 months ago

I don't think so that's would be valid. I raised a similar query here but the developers discourage inferring any directionality using LFC values. Besides, please read the paragraph from their paper below:

Next, we note that scTenifoldKnk is designed to predict DR genes rather than DE genes. DR genes might be differentially expressed upon the gene KO. To examine the expression-level changes of virtual KO perturbed genes, we performed a systematic comparison between the scTenifoldKnk results and the results of DE analysis across the three analyzed datasets: Trem2, Nkx2-1, and Hnf4ag. For each dataset, we started by computing the DE statistics for all genes. Specifically, we obtained the fold change (FC) of each gene’s expression in WT samples related to KO samples (WT/KO) using the DE analysis package MAST.51 Then, we compared FC between significant DR genes and non-significant DR genes. We found that significant DR genes, or perturbed genes predicted by scTenifoldKnk, tend to have a larger FC value (p-value < 0.05 for all three cases [Trem2, Nkx2-1, and Hnf4ag], one-sided t-tests with log2-transformed FC values) than non-significant DR genes or non-perturbed genes (Figure S4). Thus, the expression of DR genes predicted by scTenifoldKnk is more likely to be downregulated in samples of the real KO experiments.

I also tried their new tool called GenKI but is grossly slower (see issue) than this tool and takes forever to run even on 112 CPUs. I really like the simplicity of this R package and how fast it is.